Power system dominant instability mode identification method based on deep learning

By constructing node and time step correlation matrix, combining deep learning algorithms, designing cross-entropy loss functions with space-time integrated layer and weighted cross-entropy loss function, and using graph convolution and bidirectional gating loop unit model, the spatial-temporal characteristics in the existing technology are not considered enough The problem has improved the accuracy and robustness of the identification of dominant instability mode of the power system.

CN119066561APending Publication Date: 2024-12-03NORTHEAST DIANLI UNIVERSITY +1
View PDF 0 Cites 5 Cited by

Patent Information

Application Number
CN202410993976.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-24
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

The existing method for identifying dominant instability mode of power systems based on deep learning is not considered profound enough in the space-time characteristics, and the calculation is large enough, making it difficult to accurately identify dominant instability mode in complex power grids, especially in new topology and short-term windows. And the bidirectional spatiotemporal relationship between electrical quantity is ignored.

Method used

By constructing the node correlation coefficient matrix and the time step correlation coefficient matrix, combining deep learning algorithms, designing the space-time integrated layer and the three-class cross-entropy loss function with weights, using graph convolutional neural network and bidirectional gating recurrent unit model to integrate space-time information , enhance the robustness and recognition performance of the model.

Benefits of technology

Improve the accuracy and robustness of the recognition of dominant instability mode, enhance the recognition performance of the model for variable topology and short time windows, solve the problems of large amount of calculation and insufficient consideration of spatiotemporal characteristics, and take into account the dominant instability mode through spatiotemporal characteristics. The impact of

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119066561A_ABST
    Figure CN119066561A_ABST
Patent Text Reader

Abstract

The invention discloses an electric power system dominant instability mode identification method based on deep learning. The method comprises the steps of reading voltage, phase angle, active power and reactive power data in a time window of each node of an electric power system; calculating a node correlation coefficient matrix of the power system by means of a node correlation method, and calculating a time step correlation coefficient matrix by means of a spearman method; introducing the node correlation coefficient matrix into a graph convolutional neural network model, and fusing the time step correlation coefficient matrix into a bidirectional gating cycle unit model to obtain a GCN-BiGRU model capable of deeply mining spatio-temporal information; constructing a space-time integrated layer by using a multi-head attention mechanism, mining key space and time sequence features processed by GCN and Bi-GRU, and fusing space-time information in a mode of combining dimension reduction, remodeling and splicing; a cross entropy function with a weight is constructed as a loss function in training, and the problem that the number of various categories of dominant instability mode samples is unbalanced is solved; and optimizing the deep neural network model according to the final loss and carrying out dominant instability mode identification.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of power system safety and stability, and in particular to a deep learning method for identifying dominant instability modes of a power system. Background Art

[0002] Under the background of my country's "dual carbon" goal, the scale of wind power and photovoltaic grid connection has exploded, high-voltage direct current transmission has been widely used, and the power system has gradually evolved into a new power system with a high proportion of new energy and a high proportion of power electronic equipment. [1] Transient voltage instability and transient power angle instability are two typical instability forms of new power systems. Due to the mutual coupling of electrical quantities in the power system during the dynamic process, voltage instability and power angle instability often occur at the same time. [2] Therefore, accurate and rapid identification of the dominant instability mode (DIM) of the system plays a vital role in the safe and stable operation of the power grid.

[0003] At present, most scholars mainly use traditional mechanism analysis and artificial intelligence methods to identify the dominant instability mode. [3] The dominant instability mode is identified based on the characteristics of the oscillation center and the low voltage at the load center, but this method has a large amount of calculation. [4] By analyzing the coupling characteristics of power angle and voltage under different network structures, a quantitative index to characterize the degree of coupling between voltage and power angle instability is proposed. However, it is still difficult to accurately identify the dominant instability mode in scenarios where the coupling between the two is strong. It can be seen that the mechanism analysis method is limited by problems such as computational complexity and scenario universality, and the recognition accuracy and speed of relevant theoretical judgments and practical theorems in actual complex power grids are difficult to guarantee. The deep learning method is one of the important branches of artificial intelligence. Its deep neural network can realize deep mining of massive data and has the ability to self-extract, classify or regress massive data. Some scholars have applied deep learning methods to DIM, such as convolutional neural network (CNN), virtual adversarial training (VAT), and graph attention network (GAT). In order to effectively capture the key information in the input features, the literature [5] The squeeze excitation network module is introduced into the convolutional neural network, which effectively improves the prediction effect of CNN by adaptively adjusting the weights of feature channels. However, this method has limited description of the topological connection relationship of the power grid and lacks the ability to generalize applications under new topologies. [6]A semi-supervised learning framework of virtual adversarial training and mean teacher (MT) model is established for sample generation and dominant instability mode identification. Due to the small number of input samples, the evaluation model is prone to overfitting risks and has limited generalization ability under variable topology. [7] A GAT model that takes into account the correlation of node information in the graph structure and sample data is established, and non-Euclidean data is converted into Euclidean data for analysis and training, but the temporal relationship of electrical quantities is not considered. From the above analysis, it can be seen that the existing dominant instability mode identification method based on deep learning only considers the connection relationship between the nodes in the system, ignoring the evolution law of the electrical quantities of each node at different times. In addition, the identification of the dominant instability mode is a multi-space-time dimensional problem, including both forward and reverse space-time dimensions. The existing method is difficult to effectively consider the impact of the bidirectional space-time relationship of electrical quantities on the identification of the dominant instability mode.

[0004] References

[0005] [1] Zhang Zhigang, Kang Chongqing. Challenges and prospects of building a new power system under the goal of carbon neutrality[J]. Proceedings of the CSEE, 2022, 42(08): 2806-2819.

[0006] [2] Li Zonghan, Sun Huadong, Zheng Chao, et al. Identification of dominant instability forms based on reactive power distribution characteristics of wide-area branch response [J]. Proceedings of the CSEE, 2023, 43(13): 4883-4899.

[0007] [3] Hou Junxian, Han Minxiao, Dong Yifeng, et al. Judgment method of transient power angle instability and voltage instability based on oscillation center and voltage shape [J]. Power System Technology, 2016, 40(5): 1509-1515.

[0008] [4] Li Xiao, Liu Chongru, Xin Shujun, et al. Coupling mechanism analysis and coupling strength evaluation index of transient power angle stability and transient voltage stability [J]. Proceedings of the CSEE, 2021, 41(15): 5091-5107.

[0009] [5] Shi Zhongtuo, Yao Wei, Huang Yanhao, et al. Intelligent identification of dominant instability modes of power systems based on SE-CNN and simulation data[J]. Proceedings of the CSEE, 2022, 42(21): 7719-7731.

[0010] [6] Zhang Runfeng, Yao Wei, Shi Zhongtuo, et al. Semi-supervised learning framework for dominant instability pattern recognition integrating virtual adversarial training and mean teacher model[J]. Proceedings of the CSEE, 2022, 42(20): 7497-7509.

[0011] [7] Zhang R, Yao W, Shi Z, et al.A graph attention networks-based model to distinguish the transient rotor angle instability and short-term voltageinstability in power systems[J]. International Journal of Electrical Power andEnergy Systems, 2022, 137. Summary of the invention

[0012] The present invention provides a method for identifying the dominant instability mode of a power system based on deep learning, constructs a node correlation coefficient matrix and a time step correlation coefficient matrix, combines a deep learning algorithm to deeply mine spatial and bidirectional time series information, constructs a time-space integration layer and a weighted three-classification cross entropy loss function, solves the problem that the time-space characteristics of the existing model are not considered deeply enough, improves the accuracy and robustness of the dominant instability mode identification, and solves the problem of fewer instability samples in actual engineering. See the following description for details:

[0013] A method for identifying dominant instability modes of power systems based on deep learning, comprising:

[0014] S1, reads the voltage, phase angle, active and reactive power data within the time window of each node in the power system;

[0015] S2, calculate the node correlation coefficient matrix of the power system with the help of the node correlation method, and calculate the time step correlation coefficient matrix using the spearman method;

[0016] First, the node correlation coefficient matrix of the power system is calculated, and the four characteristic quantities are used as the basis for judging the correlation between nodes to calculate the similarity between nodes i and j:

[0017]

[0018] Where n is the number of feature quantities; X ik and Y ik are the kth feature value corresponding to nodes i and j respectively; X i and Y j are the means of all feature quantities of nodes i and j respectively; r ij is the correlation coefficient between nodes i and j;

[0019] Arrange the correlation coefficients between nodes according to the node positions to obtain a new node correlation coefficient matrix r:

[0020]

[0021] Then, the time step correlation coefficient matrix is ​​calculated: Spearman is used to calculate the correlation of each time step of electrical quantities. The sliding window length is l, which is the same as the design in the GCN module. The four electrical quantities of voltage, phase angle, active power and reactive power are also used.

[0022] Gas volume, time step correlation coefficient Calculation process:

[0023]

[0024] In the formula, i is the i-th feature quantity; the rank of time step t-j+1 is For any two time steps t-i+1 and t-j+1 in the sliding window, where i, j range from 1 to l;

[0025] Use the average method to aggregate the correlation matrix of the four feature quantities to obtain an average correlation matrix C t :

[0026]

[0027] Using C t As the weight matrix, c ij is the element in the weight matrix, for the hidden state h t-i+1 ,...,h t Perform weighted summation to obtain a new vector v t :

[0028]

[0029] The eigenvector matrix x t With v t Multiply element by element to get the new eigenvector x′ t :

[0030] x′ t =x t ·v t

[0031] S3, introduce the node correlation coefficient matrix into the graph convolutional neural network model, and integrate the time step correlation coefficient matrix into the bidirectional gated recurrent unit model to obtain the GCN-BiGRU model that can deeply mine spatiotemporal information;

[0032] S4, uses a multi-head attention mechanism to mine the key spatial and temporal features after GCN and Bi-GRU processing, and integrates spatiotemporal information and takes into account spatiotemporal features by combining dimensionality reduction, reshaping and splicing;

[0033] S5, construct a weighted cross entropy function as the loss function in training to solve the problem of unbalanced number of samples of each category in the dominant instability mode;

[0034] Construct a weighted cross entropy function to form the loss function in training:

[0035]

[0036] In the formula, m is the number of single training samples; y k,0 ,y k,1 and k,2 is the label of sample k belonging to system stability, voltage-dominated instability, or power angle-dominated instability, which is a Boolean value; P k,0 , P k,1 and P k,2 is the probability value given by the discriminator that the sample k belongs to system stability, voltage-dominated instability, and power angle-dominated instability, and the value range is [0,1]; w0, w1, and w2 are the weights in the cases of stability, voltage-dominated, and power angle-dominated instability, respectively;

[0037] S6, optimizes the deep neural network model according to the final loss and identifies the dominant instability mode.

[0038] The beneficial effects of the technical solution provided by the present invention are:

[0039] 1. The present invention uses a node correlation calculation method to calculate the node correlation coefficient matrix of the power system, and constructs a graph convolutional network model with the help of the node correlation coefficient matrix. The correlation between different nodes can be considered on the basis of the inherent network topology connection relationship, thereby enhancing the graph convolutional network model's ability to capture information about non-directly connected nodes in the graph structure, and at the same time improving the model's performance in identifying the dominant instability mode of the system under variable topology conditions;

[0040] 2. The present invention uses the spearman method to calculate the time step correlation coefficient matrix, pays attention to the characteristics of time steps with stronger correlation (such as before and after the fault, before and after the fault removal), further explores the internal connection of electrical quantities at each moment and accurately captures the key timing characteristics, and deeply explores the bidirectional timing characteristics of transient electrical quantities with the help of a bidirectional gated cyclic unit model, thereby improving the accuracy and robustness of the model, so that the model can still maintain a high accuracy rate in a shorter time window, which is more in line with the requirements of shorter sampling windows in existing engineering practices, and solves the problem that the existing model requires more input features during the identification process and does not consider the timing characteristics of electrical quantities;

[0041] 3. The present invention designs a feature fusion method of the spatiotemporal integrated layer, which uses a multi-head attention mechanism to mine the key spatial and temporal features after GCN and Bi-GRU processing, and fuses spatiotemporal information by combining dimensionality reduction, reshaping and splicing, effectively taking into account the influence of spatiotemporal features on the dominant instability mode and avoiding data redundancy;

[0042] 4. The present invention adds weights of three dominant instability modes on the basis of the traditional two-category cross entropy loss function, and designs a three-category loss function suitable for identifying the dominant instability modes of power systems. By increasing the weights of samples with a smaller proportion, the model pays balanced attention to each type of samples, which solves the problem that the number of unstable samples in actual engineering is small, and the misjudgment of unstable samples has a greater impact on system stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 Schematic diagram of the network structure of the deep learning power system dominant instability mode identification model;

[0044] Figure 2 It is a schematic diagram of the space-time integrated layer structure;

[0045] Figure 3 Flowchart for offline model training and online application;

[0046] Figure 4 It is the topology diagram of the modified IEEE-39 node system;

[0047] Figure 5 Schematic diagram of the confusion matrix of the GCN-BiGRU model recognition results;

[0048] Figure 6 It is a schematic diagram of the identification accuracy of DIM (dominant instability mode) of each model under different sampling window lengths;

[0049] Figure 7 It is a schematic diagram of the recognition accuracy of each model DIM under different noise conditions;

[0050] Figure 8 Schematic diagram of the input data and output data distribution of each layer. DETAILED DESCRIPTION

[0051] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention are described in further detail below.

[0052] Example 1

[0053] The embodiment of the present invention discloses a method for identifying the dominant instability mode of a power system based on deep learning, such as Figure 1 As shown, including:

[0054] Step 1: Read the voltage, phase angle, active power and reactive power in the time window of each node of the power system, calculate the node correlation coefficient of the power system using the electrical quantity of each node, and calculate the time step correlation coefficient matrix in the time window to represent the correlation between each time;

[0055] 1) The dominant instability mode of the power system is affected by many factors. In order to effectively reflect the transient process of the system, the system node voltage amplitude and phase angle time series data U t ,θ t , active power and reactive power time series data P t , Q t As input features:

[0056] U t =(u1 u2 ... u l ) (1)

[0057] θ t =(θ1 θ2 ... θ l ) (2)

[0058] P t =(p1 p2 ... p l ) (3)

[0059] Q t =(q1 q2 ... q l ) (4)

[0060] Where t is the sampling interval and l is the sampling window length.

[0061] The sampling time interval of the input electrical quantity of the power system dominant instability mode identification model is expressed as t d , sampling starts at the fault start time t c The sampling end time is t f , define the sampling window length l as:

[0062]

[0063] 2) Calculate the node correlation coefficient matrix of the power system: There is a big difference between the distance between each node and the resistance and reactance parameters between nodes in the power system. Using the traditional neighborhood matrix to describe the connection relationship of the power grid nodes, only the connection relationship between each node and the first-order neighboring node can be constructed, and the dimension is n×n. In the power system, the reactance of the transmission line is much greater than the resistance, that is, the line reactance is the weight between the nodes, resulting in the relationship between the nodes is not a simple direct connection relationship, and there is a different degree of mutual influence between the characteristics of each node. The voltage phase trajectory consists of voltage amplitude and phase angle, and the voltage phasor on the complex plane can cover the main information of the node. The change law of the active power of the node reflects the change trend of the overall load of the power system to a certain extent. The voltage level of the node and the configuration of the reactive compensation equipment make the reactive power in similar types of nodes show certain similarity. Therefore, the voltage amplitude, phase angle, active and reactive power of the electrical quantity in the power system are highly correlated with the node. The embodiment of the present invention uses these four characteristic quantities as the basis for judging the correlation between each node, and calculates the similarity of the two nodes i and j. The calculation formula is shown in formula (6).

[0064]

[0065] Where n is the number of characteristic quantities (voltage, phase angle, active power and reactive power); X ik and Y ik is the value of the kth feature value corresponding to nodes i and j; X i and Y j are the means of all feature quantities of nodes i and j respectively; r ij is the correlation coefficient between nodes i and j.

[0066] The size of the node correlation coefficient can reflect the degree of correlation between two nodes. The calculated correlation coefficients between nodes are arranged according to the node positions to obtain a new node correlation coefficient matrix r:

[0067]

[0068] Where i=j is the number of nodes in the power system.

[0069] 3) Calculate the time step correlation coefficient matrix within the time window: Introduce the time step correlation matrix, integrate the characteristics of different time steps, explore the time window position with the greatest impact in the process of power system instability, accurately model the correlation within the time window, and use Bi-GRU to introduce a bidirectional structure to capture the long-term dependency information of the past and future in the time series data, further improving the model performance. Spearman is used to calculate the correlation of each time step, and the sliding window length is l, which is the same as the design in the GCN module. Here, four electrical quantities, voltage, phase angle, active power and reactive power, are also used for calculation. The time step correlation coefficient The calculation process is shown in formula (8).

[0070]

[0071] In the formula, i is the i-th feature quantity; the rank of time step t-j+1 is For any two time steps t-i+1 and t-j+1 within the sliding window (where i, j range from 1 to l).

[0072] Use the average method to aggregate the correlation matrix of the four feature quantities to obtain an average correlation matrix C t :

[0073]

[0074] Using C t As the weight matrix, c ij is the element in the weight matrix, for the hidden state h t-i+1 ,...,h t Perform weighted summation to obtain a new vector v t :

[0075]

[0076] The eigenvector matrix x t With v t Multiply element by element to get the new eigenvector x′ t :

[0077] x′ t =x t ·v t (11)

[0078] Step 2: Based on step 1, the node correlation coefficient matrix r is used to build a GCN (graph convolutional network) model, and the time step correlation coefficient matrix is ​​integrated into the Bi-GRU (bidirectional gated recurrent unit) model to further deeply explore the spatiotemporal characteristics of electrical quantities;

[0079] 1) Construct the GCN model using the node correlation coefficient matrix r: For the graph, we first need to calculate the normalized Laplace matrix in the frequency domain using formula (12) to describe the spatial feature information of the power system topology graph:

[0080] L m =I N -D -1 / 2 rD -1 / 2 =UΔU T (12)

[0081] In the formula, I N is the N-order unit matrix; r is the correlation coefficient matrix (a total of i nodes); D is the degree matrix (i dimension) of the node corresponding to r; U is the L mThe vector matrix of eigenvalue sorting; Δ is the matrix of L m The diagonal matrix composed of eigenvalues. The relationship between r and D is:

[0082]

[0083] Based on the Fourier transform and inverse transform of the graph, the graph convolution formula applicable to the power system topology graph can be derived:

[0084] x*Gg=U(U T x⊙U T g) (14)

[0085] Where x∈R N is the i-dimensional feature vector of the power system topology map; * represents graph convolution; G is the power system topology map that needs to be convolved; g is the convolution kernel; ⊙ represents the Hadamard product.

[0086] Substituting the correlation coefficient matrix r as the new adjacency matrix into the hidden state update formula of GCN, we get:

[0087] X (l+1) =σ[D -1 / 2 rD -1 / 2 X (l) W (l) ] (15)

[0088] In the formula, D = ∑r, the calculation of the degree matrix D is based on the correlation coefficient matrix r; X (l) and X (l+1) are the feature matrices of the current layer and the next layer respectively; σ is the activation function; W (l) is the weight matrix of the current layer of the model.

[0089] The embodiment of the present invention adopts GCN based on spectral domain, uses convolution theorem to define graph convolution from spectral domain and aggregates and updates node features, and each node in the graph will be updated due to changes in the features of other nodes in the global. The topological structure of the power grid is a typical complex network composed of a large number of nodes with different attributes and edges.

[0090] 2) Incorporating the time-step correlation coefficient matrix into the Bi-GRU model: The hidden layer of the Bi-GRU network consists of a single GRU. The GRU uses a gating mechanism to capture long-distance dependencies in the input sequence, solving the gradient vanishing and gradient exploding problems that occur during the training of recurrent neural networks. The GRU uses an update gate to determine the fused information added to the hidden layer, and a reset gate to determine the amount of information discarded at the current moment.

[0091] The basic components of the GRU network are the update gate and the reset gate. The update gate formula is constructed using the feature vector containing window correlation:

[0092] z t =σ(W z ·[x′ t ,h t-1 ]) (16)

[0093] Where σ is the sigmoid activation function; x′ t is the new eigenvector; h t-1 is the candidate state passed at the previous moment. Update gate z t The larger the value, the more state information of the previous moment is retained in the current state, and the shorter dependency relationship can be adapted. t The smaller it is, the more input information at the current moment is ignored, so that GRU can model longer dependency relationships and have a certain memory.

[0094] The reset gate formula is:

[0095] r t =σ(W r ·[x′ t ,h t-1 ]) (17)

[0096] Similar to the update gate, the reset gate determines whether the hidden state of the previous moment is reset at the current moment, limiting the influence of the previous hidden state on the candidate state, r t The smaller the value, the more state information of the previous moment is written into the current candidate set. The less.

[0097] The candidate state calculation formula of GRU is as follows:

[0098]

[0099] In the formula, tanh represents the hyperbolic tangent activation function; ⊙ represents the Hadamard product. In the above formulas (16) to (18), W z , W r , W h is the input weight matrix.

[0100] The hidden state h in GRU t Considering the input data and the hidden state h at the previous moment t-1 The combined calculation results are as follows:

[0101]

[0102] When the power system encounters a large disturbance, there is a close connection between the electrical quantity at a certain moment and the historical and subsequent operating states of the power system. However, the unidirectional GRU network can only extract feature information in the forward time series direction and cannot reversely consider the connection between the future state and the current state. The Bi-GRU model trains two unidirectional GRUs in opposite directions at the same time, taking into account the bidirectional information flow of the time series. It can effectively take into account the impact of historical data and future data on the current data, and use the feature vector containing window correlation to fully extract the deep time series feature information in the electrical quantity, so as to better adapt to the changes and noise in the sequence. The expression is as follows:

[0103]

[0104] In the formula, w t and v t are the output weights of the front hidden layer and the back hidden layer of the time series at time t; and are the output states of the forward and backward GRU at time t-1 respectively.

[0105] Step 3: The present invention further establishes a spatiotemporal integration layer based on the Bi-GRU and GCN constructed in step 2, and uses a multi-head attention mechanism to mine the key spatial and temporal features processed by GCN and Bi-GRU, and integrates spatiotemporal information by combining dimensionality reduction, reshaping and splicing, effectively taking into account the impact of spatiotemporal features on the dominant instability mode;

[0106] The temporal evolution law and spatial topology information of transient electrical quantities are retained to achieve deep extraction of spatiotemporal characteristics of transient electrical quantities in the power system, avoid excessive reliance of the model on a certain dimension of spatiotemporal information, and enhance the learning ability of multi-source information. Figure 2 shown.

[0107] Depend on Figure 2It can be seen that after Bi-GRU feature mining, the input data obtains a three-dimensional tensor containing the bidirectional time series features of transient electrical quantities, and after GCN feature mining, a two-dimensional tensor containing the spatial features of transient electrical quantities of each node is obtained. In order to solve the problem of key feature loss of time series evolution features and spatial topological information in the data fusion process, this paper uses a multi-head attention mechanism to obtain the attention distribution of GCN and Bi-GRU output features at different positions and different representation subspaces, increase the weight of key spatial and temporal information, and combine the weight matrix with the data by scaling the dot product, taking into account the strong correlation information between spatiotemporal data and the dominant instability mode. Further, the data processed by Bi-GRU is reduced in dimension and reshaped into two dimensions, and spliced ​​with the two-dimensional data processed by GCN to obtain two-dimensional data representing the spatiotemporal characteristics of the power system. This method can retain the key time series evolution law and spatial topological information of transient electrical quantities, realize the deep extraction of spatiotemporal features of transient electrical quantities in the power system, avoid the model's excessive reliance on a certain dimensional feature of spatiotemporal information, enhance the learning ability of multi-source information, and more comprehensively represent the spatiotemporal characteristics of electrical quantities in the power system.

[0108] This method can obtain the complete spatiotemporal characteristic information of transient electrical quantities at each node in the system and retain the dependence on timing and spatial topology, thereby more comprehensively characterizing the spatiotemporal characteristics of electrical quantities in the power system, which helps to improve the accuracy and reliability of identifying the dominant instability mode of the power system.

[0109] Step 4: Based on the identification model designed in step 3, a weighted cross entropy function is constructed as the loss function in training to solve the problem of unbalanced number of samples of the dominant instability mode in each category;

[0110] A prominent problem in the design of data-driven power system transient stability assessment systems is the imbalance in the number of samples. That is, the vast majority of samples that can actually be obtained are stable samples, and the number of unstable samples is relatively small. However, in practical applications, more attention is paid to the accurate identification of unstable samples. Therefore, this embodiment adds weights of three dominant instability modes on the basis of the traditional two-category cross entropy loss function, and designs a three-category loss function suitable for the identification of dominant instability modes of power systems. By increasing the weights of samples with a smaller proportion, the model pays balanced attention to various types of samples, which solves the problem that the number of unstable samples in actual engineering is small, and the misjudgment of unstable samples has a greater impact on system stability. Referring to the processing techniques for edge monitoring in the field of image recognition, the following weighted cross entropy function is constructed to constitute the loss function in training:

[0111]

[0112] In the formula, m is the number of single training samples; y k,0 ,y k,1 and k,2is the label of sample k belonging to system stability, voltage-dominated instability, or power angle-dominated instability, which is a Boolean value; P k,0 , P k,1 and P k,2 is the probability value given by the discriminator that the sample k belongs to system stability, voltage-dominated instability, and power angle-dominated instability, and the value range is [0,1]; w0 is the weight when the sample is system stable; w1 is the weight when the sample is voltage-dominated instability; w2 is the weight when the power angle-dominated instability. In practice, the ratio of w0, w1, and w2 can be set according to the ratio of unstable samples to stable samples in the sampled training sample set.

[0113] According to practical engineering criteria and expert experience, combined with the node voltage and generator power angle change trajectory during the entire system fault process, the dominant instability mode of the system is identified. If the transient voltage instability and power angle instability are strongly coupled with each other, in order to accurately identify the dominant instability mode, it is necessary to further apply an emergency control strategy to verify the dominant instability mode. Specifically, if the generator is removed and the system is restored to stability, it is a power angle dominated instability, and if the load is removed and the system is restored to stability, it is a voltage dominated instability.

[0114] The embodiment of the present invention is designed as follows Figure 1 The six-layer network shown in the figure is used to construct the dominant instability mode identification model of the power system.

[0115] Depend on Figure 1 It can be seen that the first layer of the proposed dominant instability mode identification model is the data input layer, which reshapes the normalized data into a three-dimensional array representing the time series, space and power system characteristics according to the spatiotemporal characteristics, and constructs a node correlation coefficient matrix reflecting the correlation of each node in the power system; the second layer is the graph convolution layer, which realizes the aggregation and extraction of the spatial characteristics of the input electrical quantity; the third layer is the bidirectional gated recurrent unit layer, which can construct the time step correlation coefficient matrix and deeply mine the bidirectional evolution law of the electrical quantity of each node at different times; the fourth layer is the spatiotemporal integration layer, which reduces the dimension of the data processed by Bi-GRU and GCN and constructs a feature matrix with spatiotemporal characteristics; the fifth layer is the fully connected layer, which nonlinearly maps the aggregated features with the system dominant instability mode categories (system stability, voltage-dominated instability, power angle-dominated instability) through the activation function; the sixth layer is the output layer, which outputs the results of the dominant instability mode identification of the power system.

[0116] The offline training and online application process of the power system dominant instability mode identification model is as follows: Figure 3As shown. During the offline training process, the model uses the Adam optimization algorithm and supervises the learning with the cross entropy loss function suitable for the dominant instability mode identification. After the model training is completed and the target accuracy is achieved, it can be used for online dominant instability mode identification. The transient electrical quantity data is read as input using the PMU measurement device within the observation window. The model will automatically identify the system state as stable, voltage-dominated instability or power angle-dominated instability. The identification results will provide guidance for the subsequent formulation of the decision table.

[0117] Step 5: Generate a large number of data sets for offline training under different failure scenarios of the modified IEEE-39 node system, and construct evaluation indicators that can verify the performance of the model: accuracy, precision, recall, F1 score and confusion matrix;

[0118] Three wind farms, two photovoltaic farms and one traditional DC are introduced into the IEEE-39 node system to simulate the characteristics of the new power system with a high proportion of new energy and power electronic equipment. The system includes 10 traditional units, 34 transmission lines and 39 nodes. The modified IEEE-39 node system topology is as follows Figure 4 shown.

[0119] Due to the complex and changeable structure of the power system with large-scale grid-connected renewable energy, four types of data sets are set in the process of sample generation: Dataset A: power electronics account for 0%; Dataset B: power electronics account for 18%; Dataset C: power electronics account for 48%; Dataset D: a line is randomly broken based on Dataset A.

[0120] 1) Dataset A: The system has no disconnection, the ratio of induction motors is set to level 3, and three-phase short-circuit grounding faults at 11 fault locations on 34 AC lines are considered under 9 power flow levels. The specific settings are shown in Table 1, and a total of 9×3×34×11=10098 samples are generated.

[0121] 2) Dataset B: Based on Dataset A, the influence of new energy and power electronic equipment such as DC is considered. The 1200MW (18%) output of traditional units in the original IEEE-39 node system is replaced by 500MW of wind power output, 300MW of photovoltaic output and 400MW of traditional high-voltage DC input. The other scenario settings are the same as Dataset A, generating a total of 9×3×34×11=10098 samples.

[0122] Table 1 Fault scenario settings

[0123]

[0124]

[0125] 3) Dataset C: Based on Dataset A, a large proportion of power electronics is considered, accounting for 48%. The 3000MW (48%) of traditional unit output in the original IEEE-39 node system is replaced by 1500MW of wind power output, 1000MW of photovoltaic output and 400MW of traditional high-voltage direct current. The other scenario settings are the same as Dataset A, generating a total of 9×3×34×11=10098 samples.

[0126] 4) Dataset D: One line is randomly disconnected in all AC lines except the line containing the transformer, forming a total of 34 topologies. Three-phase short-circuit grounding faults at 11 fault locations on 33 lines except the disconnected line under each topology are considered respectively. The other scenario settings are the same as those of dataset A, generating a total of 9×3×34×11=10098 samples.

[0127] Through the above sample generation process, a large number of fault samples representing the existing new power system with a high proportion of new energy and power electronic equipment can be generated, which can solve the problem that the existing samples do not consider new energy and power electronic equipment enough, resulting in poor practicality. In addition, the scenarios considered are richer, further verifying the generalization and robustness of the model. The number of dominant instability modes in different data sets is shown in Table 2. Each data set is divided in a ratio of 6:2:2 to obtain the training set, validation set, and test set required for the dominant instability mode identification model.

[0128] Table 2 Samples of each category in the four datasets

[0129]

[0130] In order to comprehensively evaluate the dominant instability mode identification performance, the embodiment of the present invention adopts three evaluation indicators to comprehensively judge the classification results of the model, including accuracy, precision, recall rate and F1 score.

[0131] The above evaluation indicators can be established through the binary classification confusion matrix. The indicators in the binary classification confusion matrix are shown in Table 3. TP is the number of samples correctly predicted as positive cases, FP is the number of samples incorrectly predicted as positive cases, TN is the number of samples correctly predicted as negative cases, and FN is the number of samples incorrectly predicted as negative cases. Specifically, for multi-classification problems, each class can be considered as a positive example and its confusion matrix can be calculated for binary classification.

[0132] Table 3 Indicators in the binary confusion matrix

[0133]

[0134]

[0135] Accuracy is an important indicator that reflects the effectiveness of model classification, representing the proportion of correct results identified by the model, as shown below:

[0136]

[0137] Where n correct is the number of samples correctly identified; n samples is the number of samples to be identified.

[0138] The precision index represents the proportion of samples correctly classified as positive examples among all samples classified as positive examples; the recall rate Recall is the proportion of samples correctly classified as positive examples among all actual positive examples. Precision and recall rate require specifying a positive example. For any positive example i, precision and recall rate can be calculated by the following formula:

[0139]

[0140] Precision and recall are two metrics that weigh each other. The numerical values ​​of these two indicators can determine the accuracy of the model in terms of recognition results, but there may be a large number of missed samples. The F1 score is a combination of these two indicators, which is the sum average of precision and recall:

[0141]

[0142] In multi-classification problems, Macro-Averaged can calculate the performance of each category independently and then take the average value. Compared with Micro-Averaged, which gives different weights according to the number of category samples, Macro-Averaged gives each category the same weight, which is more affected by rare categories and is more suitable for the identification of dominant instability modes, helping to gain a deeper understanding of the model's performance in each category. Under the Macro-Averaged metric, the precision, recall, and F1 scores of multi-classification problems are:

[0143]

[0144]

[0145] Where C is the number of categories, and the dominant instability mode category C is 3.

[0146] To demonstrate the recognition performance of the model in three categories, the embodiment of the present invention generates a 3×3 confusion matrix for the recognition results, which can reflect the corresponding relationship between the prediction results of the model in different categories and the true labels. The dominant instability mode recognition confusion matrix is ​​shown in Table 4, where TP AA , FP AB and FP ACThey represent the proportion of system stability samples that are identified as system stability, voltage-dominated instability, and power angle-dominated instability; FP BA TP BB and FP BC They represent the proportion of voltage-dominated instability samples that are identified as system stability, voltage-dominated instability, and power angle-dominated instability samples; FP CA , FP CB and TP CC They represent the proportions of power angle dominated instability samples identified as stable, voltage dominated instability and power angle dominated instability samples respectively.

[0147] Table 4. Confusion matrix of dominant instability modes

[0148]

[0149] Step 6: Based on the evaluation index constructed in step 5, the identification performance of the dominant instability mode identification method is comprehensively evaluated on the modified IEEE-39 node system example;

[0150] In order to comprehensively evaluate the recognition performance of the dominant instability mode recognition method, taking dataset B as an example, the evaluation performance of the proposed GCN-BiGRU model is compared with that of the traditional convolutional network CNN model, the GCN model that only considers spatial feature information, and the Bi-GRU model that only considers temporal feature information. The confusion matrix of the GCN-BiGRU model recognition result is shown in Figure 2. Figure 5 shown.

[0151] Depend on Figure 5 From the confusion matrix, it can be seen that the accuracy of each category of the proposed dominant instability mode identification model reaches more than 97.97%, and the discrimination results of stable samples and unstable samples are relatively balanced.

[0152] For the identification process of the dominant instability mode of the power system, engineering practice prefers a faster identification speed, but it is also more challenging for feature extraction and model identification performance. The sampling time window is the biggest factor affecting the identification speed. Using a smaller time window can greatly shorten the overall time of DIM identification, but at the same time, the mapping capability requirements of the identification model are also increased due to the reduction in sampling data. To this end, the dominant instability mode identification performance of GCN-BiGRU and CNN, GCN, and Bi-GRU under different sampling window lengths is further compared and analyzed. The dominant instability mode identification results under different sampling window lengths are shown in the figure. Figure 6 shown.

[0153] Depend on Figure 6It can be seen that: 1) Under different sampling time window lengths, the dominant instability mode identification accuracy of the proposed GRU-BiGRU remains above 98%, which is higher than the identification accuracy of CNN, GCN and Bi-GRU; 2) With the increase of the sampling window length, the transient information characterizing the state of the power system increases, and the dominant instability mode identification accuracy also increases accordingly; 3) When the sampling window length is 13, the dominant instability mode identification accuracy of the proposed method is above 98%, indicating that the proposed GCN-BiGRU model can achieve good identification performance with fewer input features.

[0154] In actual engineering applications, there will be measurement errors and noise interference when collecting measurement information through PMU. Anti-interference ability is a key indicator to test model performance. To verify the robustness of the proposed GCN-BiGRU model to noise, taking dataset D as an example, Gaussian white noise with signal-to-noise ratios of 20db, 30db, and 40db is added to the test set. The recognition accuracy of each model under different noise environments is as follows: Figure 7 shown.

[0155] Depend on Figure 7 It can be seen that when there is Gaussian white noise with a signal-to-noise ratio of 40db in the test set, the accuracy of the proposed GCN-BiGRU model can still be maintained at around 98.32%, far exceeding other dominant instability pattern recognition models. Furthermore, after reducing the signal-to-noise ratio to 30db and 20db, the proposed model can still maintain an accuracy of 98.02% and 97.82%, and the accuracy only decreases by 0.49% and 0.69% compared to the noise-free case, while other methods have a large decrease in accuracy after the noise increases. Therefore, the proposed GCN-BiGRU method can still maintain a high accuracy in different noise environments and has strong robustness to noise.

[0156] In order to intuitively understand the structure and similarity of the data after optimization of each layer of the model, the classification of the three types of samples in each layer is visualized with the help of t-SNE. t-SNE can map the original data to a low-dimensional space through mathematical mapping, and find new data point distributions in the low-dimensional space by considering the probability distribution relationship between each data point and other data points, so as to intuitively understand the structure and similarity of the data.

[0157] The embodiment of the present invention uses t-SNE to respectively display the sample distribution of the input features of the dominant instability mode identification and the output features of Bi-GRU, GCN, and fully connected layer, such as Figure 8 In the figure, different colors represent different instability modes.

[0158] Depend on Figure 8It can be seen that the three types of samples of the original input features are intertwined and overlapped, and there is no obvious classification feature. After the input features are aggregated by the Bi-GRU layer and the GCN layer, the samples of system stability, voltage-dominated instability, and power angle-dominated instability show obvious aggregation characteristics in the new feature space, presenting a gradually clear interface. The high-dimensional features containing spatiotemporal characteristics are further reduced and fused through the spatiotemporal integration layer, and the output features of the fully connected layer are completely separated by using the linear combination and nonlinear transformation of the fully connected layer. Both Bi-GRU and GCN deeply extract features at the temporal and spatial levels, and combine the mined spatiotemporal features through the spatiotemporal integration layer, thereby improving the recognition ability of different dominant instability modes and providing strong support for the interpretability of dominant instability mode identification. It can be seen that the GCN-BiGRU model can not only effectively capture the spatiotemporal relationship between samples, but also achieve accurate and reliable classification in high-dimensional feature space.

[0159] The embodiment of the present invention proposes a method for identifying the dominant instability mode of a power system based on a graph convolutional network and a bidirectional gated recurrent unit. Through simulation analysis of the modified IEEE-39 example, the following conclusions can be drawn:

[0160] 1) The embodiment of the present invention further explores the deep correlation between nodes and constructs a graph convolutional network model using a node correlation coefficient matrix, thereby enhancing the model's ability to capture information about non-directly connected nodes in the graph structure and improving the model's recognition performance under variable topology conditions.

[0161] 2) The embodiment of the present invention utilizes the characteristics of time steps with stronger correlation to further explore the intrinsic connection of electrical quantities at each moment and accurately capture key timing characteristics. With the help of bidirectional gated recurrent units, the bidirectional timing characteristics of transient electrical quantities are deeply explored, thereby improving the accuracy and robustness of the model.

[0162] 3) The embodiment of the present invention designs a feature fusion method of the spatiotemporal integrated layer, which uses a multi-head attention mechanism to mine the key spatial and temporal features after GCN and Bi-GRU processing, and effectively takes into account the impact of spatiotemporal features on the dominant instability mode.

[0163] 4) The present invention adds weights of three dominant instability modes on the basis of the traditional binary cross entropy loss function, and achieves balanced attention of the model to various types of samples by increasing the weights of samples with a smaller proportion.

[0164] Those skilled in the art will appreciate that the accompanying drawing is only a schematic diagram of a preferred embodiment, and the serial numbers of the embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0165] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for identifying the dominant instability mode of a power system based on deep learning, characterized in that: include: S1, reads the voltage, phase angle, active and reactive power data within the time window of each node in the power system; S2, calculate the node correlation coefficient matrix of the power system with the help of the node correlation method, and calculate the time step correlation coefficient matrix using the spearman method; First, the node correlation coefficient matrix of the power system is calculated, and the four characteristic quantities of voltage, phase angle, active power and reactive power are used as the basis for judging the correlation between each node, and the similarity between nodes i and j is calculated: Where n is the number of feature quantities; X ik and Y ik are the kth feature value corresponding to nodes i and j respectively; X i and Y j are the means of all feature quantities of nodes i and j respectively; r ij is the correlation coefficient between nodes i and j; Arrange the correlation coefficients between nodes according to the node positions to obtain a new node correlation coefficient matrix r: Then, the time step correlation coefficient matrix is ​​calculated: Spearman is used to calculate the correlation of each time step of electrical quantities. The sliding window length is l, which is the same as the design in the GCN module. The four electrical quantities of voltage, phase angle, active power and reactive power are also used. The time step correlation coefficient Calculation process: In the formula, i is the i-th feature quantity; the rank of time step t-j+1 is For any two time steps t-i+1 and t-j+1 in the sliding window, where i, j range from 1 to l; Use the average method to aggregate the correlation matrix of the four feature quantities to obtain an average correlation matrix C t : Using C t As the weight matrix, c ij is the element in the weight matrix, for the hidden state h t-i+1 ,...,h t Perform weighted summation to obtain a new vector v t : The eigenvector matrix x t With v t Multiply element by element to get the new eigenvector x t ′: x′ t =x t ·v t S3, introduce the node correlation coefficient matrix into the graph convolutional neural network model, and integrate the time step correlation coefficient matrix into the bidirectional gated recurrent unit model to obtain the GCN-BiGRU model that can deeply mine spatiotemporal information; S4, uses a multi-head attention mechanism to mine the key spatial and temporal features after GCN and Bi-GRU processing, and integrates spatiotemporal information and takes into account spatiotemporal features by combining dimensionality reduction, reshaping and splicing; S5, construct a weighted cross entropy function as the loss function in training to solve the problem of unbalanced number of samples of each category in the dominant instability mode; Construct a weighted cross entropy function to form the loss function in training: In the formula, m is the number of single training samples; y k,0 ,y k,1 and k,2 is the label of sample k belonging to system stability, voltage-dominated instability, or power angle-dominated instability, which is a Boolean value; P k,0 , P k,1 and P k,2 is the probability value given by the discriminator that the sample k belongs to system stability, voltage-dominated instability, and power angle-dominated instability, and the value range is [0,1]; w0, w1, and w2 are the weights in the cases of stability, voltage-dominated, and power angle-dominated instability, respectively; S6, optimizes the deep neural network model according to the final loss and identifies the dominant instability mode.

Citation Information

Cited By

  • Equipment monitoring index screening method and system based on time sequence modeling

    CN120011775A

  • A method and system for screening device monitoring indicators based on time series modeling

    CN120011775B

  • Dual-drive fault diagnosis method for fusion reactor vacuum pump set

    CN120175630A

  • A dual-drive fault diagnosis method for fusion reactor vacuum pump groups

    CN120175630B

  • Multi-source data transmission and edge computing integrated device and method in special scene

    CN121037912A